Awesome

Rotated Mask R-CNN

By Shijie Looi.

(Paper to be published soon...or not, depends on schedule)

This project is based on maskrcnn-benchmark. Rotation NMS layers were based on RRPN.

The Problem With MaskRCNN (and Bounding Boxes)

Due to bounding box ambiguity, Mask R-CNN fails in relatively dense scenes with objects of the same class, particularly if those objects have high bounding box overlap. In these scenes, both recall (due to NMS) and precision (foreground instance class ambiguity) are affected.

MaskRCNN takes a bounding box input to output a single foreground (instance) segmentation per class. The hidden assumption here (as is common in many detection networks) is that a good bounding box contains just one object in that class. This is not the case for dense scenes like the pencil image above.

Unfortunately, such scenes are underrepresented in the most popular instance segmentation datasets - MSCOCO, Pascal VOC, Cityscapes. Yet they are not uncommon in many real-world applications e.g. robotics/logistics, household objects i.e. pens/chopsticks, etc. As a result, I've released a simple, small dataset called PPC - Pens, Pencils, Chopsticks (see below), and show the significant difference between Mask R-CNN and Rotated Mask R-CNN in such scenes.

Rotated Mask R-CNN

Rotated Mask R-CNN resolves some of these issues by adopting a rotated bounding box representation.

This repository extends Faster R-CNN, Mask R-CNN, or even RPN-only to work with rotated bounding boxes.

This work also builds on the Mask Scoring R-CNN ('MS R-CNN') paper by learning the quality of the predicted instance masks (maskscoring_rcnn).

The repo master branch is fully merged upstream with the latest master branch of maskrcnn-benchmark (as of 25/07/2019)

Results

COCO
Trained on coco/train2014, evaluated on coco/val2014

PPC (Pens, Pencils, Chopsticks) PPC (Pens, Pencils, Chopsticks) dataset: Link
Trained on train.json, evaluated on test.json (pens & pencils only, no chopstick class)

Additional Features

• Soft NMS (Implemented for both bounding box and rotated detections. Original repo)
• Mask IoU head (From maskscoring_rcnn). This is a better refactored version than the original repo - the original does not have batch inference/testing in the Mask IoU layer.

TODO

• Keypoints for rotated bounding boxes

Install

Check INSTALL.md for installation instructions.

Prepare Data

  mkdir -p datasets/coco
  ln -s /path_to_coco_dataset/annotations datasets/coco/annotations
  ln -s /path_to_coco_dataset/train2014 datasets/coco/train2014
  ln -s /path_to_coco_dataset/test2014 datasets/coco/test2014
  ln -s /path_to_coco_dataset/val2014 datasets/coco/val2014

Configs

All example configs related to rotated maskrcnn are in configs/rotated folder

• Rotated Mask R-CNN (default): configs/rotated/e2e_mask_rcnn_R_50_FPN_1x.yaml
• Rotated Mask Scoring R-CNN (MS-RCNN, gives slightly better mask precision than default): configs/rotated/e2e_ms_rcnn_R_50_FPN_1x.yaml
• Rotated Faster R-CNN only (without mask outputs): configs/rotated/e2e_faster_rcnn_R_50_C4_1x.yaml

Pretrained Models

Pre-trained models (and config) on MSCOCO can be found here:

• Rotated MS R-CNN
• MS R-CNN

Training

Single GPU Training (default training on MSCOCO 2014)

  python tools/train_net.py --config-file "configs/rotated/e2e_ms_rcnn_R_50_FPN_1x.yaml" SOLVER.IMS_PER_BATCH 2 SOLVER.BASE_LR 0.0025 SOLVER.MAX_ITER 720000 SOLVER.STEPS "(480000, 640000)" TEST.IMS_PER_BATCH 1

Multi-GPU Training (default training on MSCOCO 2014)

  export NGPUS=8
  python -m torch.distributed.launch --nproc_per_node=$NGPUS tools/train_net.py --config-file "configs/rotated/e2e_ms_rcnn_R_50_FPN_1x.yaml" 

For more details, see README.md in https://github.com/facebookresearch/maskrcnn-benchmark

Training Your Own Dataset

• Step 1: Convert your own dataset to COCO annotation format (a json file). I've created one very simple repo convert_to_coco to do this; examples include ICDAR datasets to COCO.
• Step 2: Add the path of the dataset image directory and coco-format annotation file (from Step 1) into maskrcnn_benchmark/config/paths_catalog.py
• Step 3: In your .yaml config file (e.g. "configs/rotated/e2e_ms_rcnn_R_50_FPN_1x.yaml"), change the DATASETS.TRAIN value to the stuff you added in paths_catalog.py. DATASETS.TEST is optional
  NOTE: SOLVER.MAX_ITER is the default training iterations used for COCO. You'll want to change this to roughly N images in your dataset, multiplied by 5-10x, divided by GPUs used for training. E.g. Set MAX_ITER to 5000-10000 if you have 1000 images on 1 GPU. Make sure to also adjust the learning rate accordingly.
  I would strongly suggest reading "2. Modify the cfg parameters" in https://github.com/facebookresearch/maskrcnn-benchmark/blob/master/README.md to make sure that your training is properly optimized.
• Step 4: Run python tools/train_net.py --config-file myconfig_in_step3.yaml

Testing

  python tools/test_net.py --config-file "configs/rotated/e2e_ms_rcnn_R_50_FPN_1x.yaml" --ckpt checkpoints/rotated/mscoco_msrcnn/model_final.pth  

For more details, see README.md in https://github.com/facebookresearch/maskrcnn-benchmark

Note that detection ("bbox") results are not relevant to Rotated Mask R-CNN, since detections are defined as bounding boxes, while Rotated Mask R-CNN outputs rotated bounding boxes.

Inference

  python my_tools/infer_demo.py

Be sure to change the input values e.g. config_file (.yaml), model_file (.pth), image_dir

Performance

• Memory: Almost identical to Mask RCNN (with just a few more parameters)
• Speed: slightly slower (~10%) during inference, 30-50% slower during training

Visualizing Rotated RPN Anchors

  python my_tools/vis_rpn_anchors.py

Can be a useful tool for visualizing base Rotated RPN anchors. Use it to adjust the anchor sizes and ratios (and angles, if needed) for your application.

Other Examples

Acknowledgment

The work was done at Dorabot Inc.

Citations

If you find Rotated Mask R-CNN useful in your research, please consider citing:

@misc{looi2019rotatedmrcnn,
  author = {Shijie Looi},
  title = {Rotated Mask R-CNN: From Bounding Boxes To Rotated Bounding Boxes},
  year = {2019},
  publisher = {GitHub},
  howpublished = {\url{https://github.com/mrlooi/rotated_maskrcnn}}
}

Note that this is not a standard BibTeX citation.

Digital Object Identifier (DOI)

License

rotated_maskrcnn is released under the MIT license. See LICENSE for additional details.

Thanks to the Third Party Libs

maskrcnn-benchmark
Pytorch